Automated Computational Inference Engine for Bayesian Source Reconstruction: Application to Some Detections/Non-detections Made in the CTBT International Monitoring System

We address the inverse problem of source reconstruction from a Bayesian perspective. Bayesian probability theory provides a natural and logically consistent (principled) approach for source reconstruction. To facilitate Bayesian source reconstruction, we develop an automated computational inference engine for efficient sampling from the potentially complex source parameter (or hypothesis) space. The engine explores this hypothesis space by utilizing “guided” importance sampling from a sequence of intermediate distributions that smoothly connect (bridge) the prior and the posterior distributions, with the sampling from each of the bridging distributions conducted using a general multiple-try differential evolution adaptive Metropolis algorithm. Furthermore, we show how to incorporate rigorously the available sensor information about detections and non-detections into the likelihood function (one of the two key input quantities for the computational inference engine). The performance of the automated computational inference engine is illustrated with an application to (re-analysis of) a very small (challenging) activity concentration data set consisting of detections and non-detections made by two radionuclide observation stations in the International Monitoring System.